Two weeks ago I’ve participated in the conference “3D Data Processing, Visualization and Transmission” that took place in Paris. Among many papers and posters related to problems of 3D vision I found some very interesting works that could be a matter of interest both for myself and our vision group in Göttingen. Today I’ve given a short overview about five most interesting papers for my group.
 Derek Hoiem et.al. ,”Recovering Occlusion Boundaries from a Single Image” [link]
A part of this paper was presented by one of the invited speakers, Alexei Efros. The goal of the current study is to recover the occlusion boundaries and depth ordering from a single image. Single images, being merely a projection of the 3D scene, contain significant information about occlusions. Results for outdoors images look quite nice. However, using only one camera no evaluations about distances between objects and camera can be done. Also some “ground truths” look strange representing some different objects as a single segment. For some examples segmentation results look even better than a ground truth. By the way, it’s an open source project.
 Mikhail Sizintsev et.al. ,”GPU Accelerated Realtime Stereo for Augmented Reality” [link]
This paper presents a real-time, robust and quite accurate stereo matching algorithm based on a coarse-to-fine architecture. It’s based on the block-based method where image pyramids are being used in order to perform stereo estimation progressively from coarser to finer levels of the pyramid. Also they use some smart and efficient refinement techniques to fix erroneous disparities on the next level. The algorithm is implemented on GPU and runs in real-time. On the conference I have played a bit with the proposed system and I had an impression that it provides even better dense disparities for weakly textured objects than phase-based stereo algorithms. However, it’s not an open source project, since the algorithm was developed for US Air Force. If you wanna use this stereo method – just reimplement it 😉
 Julien Michot et.al. ,”Bi-Objective Bundle Adjustment With Application to Multi-Sensor SLAM” [link]
Here multi-sensor data fusion problem in a Simultaneous Localization And Mapping (SLAM) application is considered. As sensors they used an odometer and a gyroscope. I’m not familiar with approaches used in SLAM applications, but in this paper they are claiming that Bundle Adjustment (BA) has better results than Extended Kalman Filter (EKF). BA is slower than EKF but it is still real-time. I found this paper quite interesting, however, the experimental results are poor, so from the paper it’s not clear how efficient and robust the method is.
 Henrik Aanæs et.al. ,”On Recall Rate of Interest Point Detectors” [link]
In this study authors provide a method for evaluating interest point detectors independently of image descriptors. This paper got the best paper award. The main contribution is an investigation of nine established interest point detectors, which provides new insights to the stability of these detectors with respect to large changes in viewpoint and scale. There are no many new scientific ideas in this paper, but I liked a lot their data collection setup and quantitative analysis in the Experiments section.
 Andrea Albarelli et.al. ,”Robust Game-Theoretic Inlier Selection for Bundle Adjustment” [link]
The last paper is also related to feature descriptor algorithms. It got the best student paper award. Once the winner was announced on the banquet, authors have already left the conference;-) A novel game-theoretic technique is introduced that performs an accurate feature matching between multiple views of the same object. as a preliminary step for bundle adjustment (BA). I liked the matching strategy that exploits Game Theory giving payoffs to players. Also the results section and quantitative analysis are worth in this paper. I liked it.